git.oblomov.eu Git - wine/blob - dlls/wined3d/ati_fragment_shader.c

   1 /*
   2  * Fixed function pipeline replacement using GL_ATI_fragment_shader
   3  *
   4  * Copyright 2008 Stefan Dösinger(for CodeWeavers)
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  19  */
  20
  21 #include "config.h"
  22
  23 #include <math.h>
  24 #include <stdio.h>
  25
  26 #include "wined3d_private.h"
  27
  28 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  29 WINE_DECLARE_DEBUG_CHANNEL(d3d);
  30
  31 /* GL locking for state handlers is done by the caller. */
  32
  33 /* Some private defines, Constant associations, etc.
  34  * Env bump matrix and per stage constant should be independent,
  35  * a stage that bump maps can't read the per state constant
  36  */
  37 #define ATI_FFP_CONST_BUMPMAT(i) (GL_CON_0_ATI + i)
  38 #define ATI_FFP_CONST_CONSTANT0 GL_CON_0_ATI
  39 #define ATI_FFP_CONST_CONSTANT1 GL_CON_1_ATI
  40 #define ATI_FFP_CONST_CONSTANT2 GL_CON_2_ATI
  41 #define ATI_FFP_CONST_CONSTANT3 GL_CON_3_ATI
  42 #define ATI_FFP_CONST_CONSTANT4 GL_CON_4_ATI
  43 #define ATI_FFP_CONST_CONSTANT5 GL_CON_5_ATI
  44 #define ATI_FFP_CONST_TFACTOR   GL_CON_6_ATI
  45
  46 /* GL_ATI_fragment_shader specific fixed function pipeline description. "Inherits" from the common one */
  47 struct atifs_ffp_desc
  48 {
  49     struct ffp_frag_desc parent;
  50     GLuint shader;
  51     unsigned int num_textures_used;
  52 };
  53
  54 struct atifs_private_data
  55 {
  56     struct wine_rb_tree fragment_shaders; /* A rb-tree to track fragment pipeline replacement shaders */
  57 };
  58
  59 static const char *debug_dstmod(GLuint mod) {
  60     switch(mod) {
  61         case GL_NONE:               return "GL_NONE";
  62         case GL_2X_BIT_ATI:         return "GL_2X_BIT_ATI";
  63         case GL_4X_BIT_ATI:         return "GL_4X_BIT_ATI";
  64         case GL_8X_BIT_ATI:         return "GL_8X_BIT_ATI";
  65         case GL_HALF_BIT_ATI:       return "GL_HALF_BIT_ATI";
  66         case GL_QUARTER_BIT_ATI:    return "GL_QUARTER_BIT_ATI";
  67         case GL_EIGHTH_BIT_ATI:     return "GL_EIGHTH_BIT_ATI";
  68         case GL_SATURATE_BIT_ATI:   return "GL_SATURATE_BIT_ATI";
  69         default:                    return "Unexpected modifier\n";
  70     }
  71 }
  72
  73 static const char *debug_argmod(GLuint mod) {
  74     switch(mod) {
  75         case GL_NONE:
  76             return "GL_NONE";
  77
  78         case GL_2X_BIT_ATI:
  79             return "GL_2X_BIT_ATI";
  80         case GL_COMP_BIT_ATI:
  81             return "GL_COMP_BIT_ATI";
  82         case GL_NEGATE_BIT_ATI:
  83             return "GL_NEGATE_BIT_ATI";
  84         case GL_BIAS_BIT_ATI:
  85             return "GL_BIAS_BIT_ATI";
  86
  87         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI:
  88             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI";
  89         case GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI:
  90             return "GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI";
  91         case GL_2X_BIT_ATI | GL_BIAS_BIT_ATI:
  92             return "GL_2X_BIT_ATI | GL_BIAS_BIT_ATI";
  93         case GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI:
  94             return "GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI";
  95         case GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI:
  96             return "GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI";
  97         case GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
  98             return "GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
  99
 100         case GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 101             return "GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 102         case GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 103             return "GL_2X_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 104         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI:
 105             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_BIAS_BIT_ATI";
 106         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI:
 107             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI";
 108
 109         case GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI:
 110             return "GL_2X_BIT_ATI | GL_COMP_BIT_ATI | GL_NEGATE_BIT_ATI | GL_BIAS_BIT_ATI";
 111
 112         default:
 113             return "Unexpected argmod combination\n";
 114     }
 115 }
 116 static const char *debug_register(GLuint reg) {
 117     switch(reg) {
 118         case GL_REG_0_ATI:                  return "GL_REG_0_ATI";
 119         case GL_REG_1_ATI:                  return "GL_REG_1_ATI";
 120         case GL_REG_2_ATI:                  return "GL_REG_2_ATI";
 121         case GL_REG_3_ATI:                  return "GL_REG_3_ATI";
 122         case GL_REG_4_ATI:                  return "GL_REG_4_ATI";
 123         case GL_REG_5_ATI:                  return "GL_REG_5_ATI";
 124
 125         case GL_CON_0_ATI:                  return "GL_CON_0_ATI";
 126         case GL_CON_1_ATI:                  return "GL_CON_1_ATI";
 127         case GL_CON_2_ATI:                  return "GL_CON_2_ATI";
 128         case GL_CON_3_ATI:                  return "GL_CON_3_ATI";
 129         case GL_CON_4_ATI:                  return "GL_CON_4_ATI";
 130         case GL_CON_5_ATI:                  return "GL_CON_5_ATI";
 131         case GL_CON_6_ATI:                  return "GL_CON_6_ATI";
 132         case GL_CON_7_ATI:                  return "GL_CON_7_ATI";
 133
 134         case GL_ZERO:                       return "GL_ZERO";
 135         case GL_ONE:                        return "GL_ONE";
 136         case GL_PRIMARY_COLOR:              return "GL_PRIMARY_COLOR";
 137         case GL_SECONDARY_INTERPOLATOR_ATI: return "GL_SECONDARY_INTERPOLATOR_ATI";
 138
 139         default:                            return "Unknown register\n";
 140     }
 141 }
 142
 143 static const char *debug_swizzle(GLuint swizzle) {
 144     switch(swizzle) {
 145         case GL_SWIZZLE_STR_ATI:        return "GL_SWIZZLE_STR_ATI";
 146         case GL_SWIZZLE_STQ_ATI:        return "GL_SWIZZLE_STQ_ATI";
 147         case GL_SWIZZLE_STR_DR_ATI:     return "GL_SWIZZLE_STR_DR_ATI";
 148         case GL_SWIZZLE_STQ_DQ_ATI:     return "GL_SWIZZLE_STQ_DQ_ATI";
 149         default:                        return "unknown swizzle";
 150     }
 151 }
 152
 153 static const char *debug_rep(GLuint rep) {
 154     switch(rep) {
 155         case GL_NONE:                   return "GL_NONE";
 156         case GL_RED:                    return "GL_RED";
 157         case GL_GREEN:                  return "GL_GREEN";
 158         case GL_BLUE:                   return "GL_BLUE";
 159         default:                        return "unknown argrep";
 160     }
 161 }
 162
 163 static const char *debug_op(GLuint op) {
 164     switch(op) {
 165         case GL_MOV_ATI:                return "GL_MOV_ATI";
 166         case GL_ADD_ATI:                return "GL_ADD_ATI";
 167         case GL_MUL_ATI:                return "GL_MUL_ATI";
 168         case GL_SUB_ATI:                return "GL_SUB_ATI";
 169         case GL_DOT3_ATI:               return "GL_DOT3_ATI";
 170         case GL_DOT4_ATI:               return "GL_DOT4_ATI";
 171         case GL_MAD_ATI:                return "GL_MAD_ATI";
 172         case GL_LERP_ATI:               return "GL_LERP_ATI";
 173         case GL_CND_ATI:                return "GL_CND_ATI";
 174         case GL_CND0_ATI:               return "GL_CND0_ATI";
 175         case GL_DOT2_ADD_ATI:           return "GL_DOT2_ADD_ATI";
 176         default:                        return "unexpected op";
 177     }
 178 }
 179
 180 static const char *debug_mask(GLuint mask) {
 181     switch(mask) {
 182         case GL_NONE:                           return "GL_NONE";
 183         case GL_RED_BIT_ATI:                    return "GL_RED_BIT_ATI";
 184         case GL_GREEN_BIT_ATI:                  return "GL_GREEN_BIT_ATI";
 185         case GL_BLUE_BIT_ATI:                   return "GL_BLUE_BIT_ATI";
 186         case GL_RED_BIT_ATI | GL_GREEN_BIT_ATI: return "GL_RED_BIT_ATI | GL_GREEN_BIT_ATI";
 187         case GL_RED_BIT_ATI | GL_BLUE_BIT_ATI:  return "GL_RED_BIT_ATI | GL_BLUE_BIT_ATI";
 188         case GL_GREEN_BIT_ATI | GL_BLUE_BIT_ATI:return "GL_GREEN_BIT_ATI | GL_BLUE_BIT_ATI";
 189         case GL_RED_BIT_ATI | GL_GREEN_BIT_ATI | GL_BLUE_BIT_ATI:return "GL_RED_BIT_ATI | GL_GREEN_BIT_ATI | GL_BLUE_BIT_ATI";
 190         default:                                return "Unexpected writemask";
 191     }
 192 }
 193
 194 static void wrap_op1(const struct wined3d_gl_info *gl_info, GLuint op, GLuint dst, GLuint dstMask, GLuint dstMod,
 195         GLuint arg1, GLuint arg1Rep, GLuint arg1Mod)
 196 {
 197     if(dstMask == GL_ALPHA) {
 198         TRACE("glAlphaFragmentOp1ATI(%s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst), debug_dstmod(dstMod),
 199               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod));
 200         GL_EXTCALL(glAlphaFragmentOp1ATI(op, dst, dstMod, arg1, arg1Rep, arg1Mod));
 201     } else {
 202         TRACE("glColorFragmentOp1ATI(%s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst),
 203               debug_mask(dstMask), debug_dstmod(dstMod),
 204               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod));
 205         GL_EXTCALL(glColorFragmentOp1ATI(op, dst, dstMask, dstMod, arg1, arg1Rep, arg1Mod));
 206     }
 207 }
 208
 209 static void wrap_op2(const struct wined3d_gl_info *gl_info, GLuint op, GLuint dst, GLuint dstMask, GLuint dstMod,
 210         GLuint arg1, GLuint arg1Rep, GLuint arg1Mod, GLuint arg2, GLuint arg2Rep, GLuint arg2Mod)
 211 {
 212     if(dstMask == GL_ALPHA) {
 213         TRACE("glAlphaFragmentOp2ATI(%s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst), debug_dstmod(dstMod),
 214               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 215               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod));
 216         GL_EXTCALL(glAlphaFragmentOp2ATI(op, dst, dstMod, arg1, arg1Rep, arg1Mod, arg2, arg2Rep, arg2Mod));
 217     } else {
 218         TRACE("glColorFragmentOp2ATI(%s, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst),
 219               debug_mask(dstMask), debug_dstmod(dstMod),
 220               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 221               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod));
 222         GL_EXTCALL(glColorFragmentOp2ATI(op, dst, dstMask, dstMod, arg1, arg1Rep, arg1Mod, arg2, arg2Rep, arg2Mod));
 223     }
 224 }
 225
 226 static void wrap_op3(const struct wined3d_gl_info *gl_info, GLuint op, GLuint dst, GLuint dstMask, GLuint dstMod,
 227         GLuint arg1, GLuint arg1Rep, GLuint arg1Mod, GLuint arg2, GLuint arg2Rep, GLuint arg2Mod,
 228         GLuint arg3, GLuint arg3Rep, GLuint arg3Mod)
 229 {
 230     if(dstMask == GL_ALPHA) {
 231         /* Leave some free space to fit "GL_NONE, " in to align most alpha and color op lines */
 232         TRACE("glAlphaFragmentOp3ATI(%s, %s,          %s, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst), debug_dstmod(dstMod),
 233               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 234               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod),
 235               debug_register(arg3), debug_rep(arg3Rep), debug_argmod(arg3Mod));
 236         GL_EXTCALL(glAlphaFragmentOp3ATI(op, dst, dstMod,
 237                                          arg1, arg1Rep, arg1Mod,
 238                                          arg2, arg2Rep, arg2Mod,
 239                                          arg3, arg3Rep, arg3Mod));
 240     } else {
 241         TRACE("glColorFragmentOp3ATI(%s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s, %s)\n", debug_op(op), debug_register(dst),
 242               debug_mask(dstMask), debug_dstmod(dstMod),
 243               debug_register(arg1), debug_rep(arg1Rep), debug_argmod(arg1Mod),
 244               debug_register(arg2), debug_rep(arg2Rep), debug_argmod(arg2Mod),
 245               debug_register(arg3), debug_rep(arg3Rep), debug_argmod(arg3Mod));
 246         GL_EXTCALL(glColorFragmentOp3ATI(op, dst, dstMask, dstMod,
 247                                          arg1, arg1Rep, arg1Mod,
 248                                          arg2, arg2Rep, arg2Mod,
 249                                          arg3, arg3Rep, arg3Mod));
 250     }
 251 }
 252
 253 static GLuint register_for_arg(DWORD arg, const struct wined3d_gl_info *gl_info,
 254         unsigned int stage, GLuint *mod, GLuint *rep, GLuint tmparg)
 255 {
 256     GLenum ret;
 257
 258     if(mod) *mod = GL_NONE;
 259     if(arg == ARG_UNUSED)
 260     {
 261         if (rep) *rep = GL_NONE;
 262         return -1; /* This is the marker for unused registers */
 263     }
 264
 265     switch(arg & WINED3DTA_SELECTMASK) {
 266         case WINED3DTA_DIFFUSE:
 267             ret = GL_PRIMARY_COLOR;
 268             break;
 269
 270         case WINED3DTA_CURRENT:
 271             /* Note that using GL_REG_0_ATI for the passed on register is safe because
 272              * texture0 is read at stage0, so in the worst case it is read in the
 273              * instruction writing to reg0. Afterwards texture0 is not used any longer.
 274              * If we're reading from current
 275              */
 276             ret = stage ? GL_REG_0_ATI : GL_PRIMARY_COLOR;
 277             break;
 278
 279         case WINED3DTA_TEXTURE:
 280             ret = GL_REG_0_ATI + stage;
 281             break;
 282
 283         case WINED3DTA_TFACTOR:
 284             ret = ATI_FFP_CONST_TFACTOR;
 285             break;
 286
 287         case WINED3DTA_SPECULAR:
 288             ret = GL_SECONDARY_INTERPOLATOR_ATI;
 289             break;
 290
 291         case WINED3DTA_TEMP:
 292             ret = tmparg;
 293             break;
 294
 295         case WINED3DTA_CONSTANT:
 296             FIXME("Unhandled source argument WINED3DTA_TEMP\n");
 297             ret = GL_CON_0_ATI;
 298             break;
 299
 300         default:
 301             FIXME("Unknown source argument %d\n", arg);
 302             ret = GL_ZERO;
 303     }
 304
 305     if(arg & WINED3DTA_COMPLEMENT) {
 306         if(mod) *mod |= GL_COMP_BIT_ATI;
 307     }
 308     if(arg & WINED3DTA_ALPHAREPLICATE) {
 309         if(rep) *rep = GL_ALPHA;
 310     } else {
 311         if(rep) *rep = GL_NONE;
 312     }
 313     return ret;
 314 }
 315
 316 static GLuint find_tmpreg(const struct texture_stage_op op[MAX_TEXTURES])
 317 {
 318     int lowest_read = -1;
 319     int lowest_write = -1;
 320     int i;
 321     BOOL tex_used[MAX_TEXTURES];
 322
 323     memset(tex_used, 0, sizeof(tex_used));
 324     for(i = 0; i < MAX_TEXTURES; i++) {
 325         if(op[i].cop == WINED3DTOP_DISABLE) {
 326             break;
 327         }
 328
 329         if(lowest_read == -1 &&
 330           (op[i].carg1 == WINED3DTA_TEMP || op[i].carg2 == WINED3DTA_TEMP || op[i].carg0 == WINED3DTA_TEMP ||
 331            op[i].aarg1 == WINED3DTA_TEMP || op[i].aarg2 == WINED3DTA_TEMP || op[i].aarg0 == WINED3DTA_TEMP)) {
 332             lowest_read = i;
 333         }
 334
 335         if(lowest_write == -1 && op[i].dst == tempreg) {
 336             lowest_write = i;
 337         }
 338
 339         if(op[i].carg1 == WINED3DTA_TEXTURE || op[i].carg2 == WINED3DTA_TEXTURE || op[i].carg0 == WINED3DTA_TEXTURE ||
 340            op[i].aarg1 == WINED3DTA_TEXTURE || op[i].aarg2 == WINED3DTA_TEXTURE || op[i].aarg0 == WINED3DTA_TEXTURE) {
 341             tex_used[i] = TRUE;
 342         }
 343     }
 344
 345     /* Temp reg not read? We don't need it, return GL_NONE */
 346     if(lowest_read == -1) return GL_NONE;
 347
 348     if(lowest_write >= lowest_read) {
 349         FIXME("Temp register read before being written\n");
 350     }
 351
 352     if(lowest_write == -1) {
 353         /* This needs a test. Maybe we are supposed to return 0.0/0.0/0.0/0.0, or fail drawprim, or whatever */
 354         FIXME("Temp register read without being written\n");
 355         return GL_REG_1_ATI;
 356     } else if(lowest_write >= 1) {
 357         /* If we're writing to the temp reg at earliest in stage 1, we can use register 1 for the temp result.
 358          * there may be texture data stored in reg 1, but we do not need it any longer since stage 1 already
 359          * read it
 360          */
 361         return GL_REG_1_ATI;
 362     } else {
 363         /* Search for a free texture register. We have 6 registers available. GL_REG_0_ATI is already used
 364          * for the regular result
 365          */
 366         for(i = 1; i < 6; i++) {
 367             if(!tex_used[i]) {
 368                 return GL_REG_0_ATI + i;
 369             }
 370         }
 371         /* What to do here? Report it in ValidateDevice? */
 372         FIXME("Could not find a register for the temporary register\n");
 373         return 0;
 374     }
 375 }
 376
 377 static GLuint gen_ati_shader(const struct texture_stage_op op[MAX_TEXTURES], const struct wined3d_gl_info *gl_info)
 378 {
 379     GLuint ret = GL_EXTCALL(glGenFragmentShadersATI(1));
 380     unsigned int stage;
 381     GLuint arg0, arg1, arg2, extrarg;
 382     GLuint dstmod, argmod0, argmod1, argmod2, argmodextra;
 383     GLuint rep0, rep1, rep2;
 384     GLuint swizzle;
 385     GLuint tmparg = find_tmpreg(op);
 386     GLuint dstreg;
 387
 388     if(!ret) {
 389         ERR("Failed to generate a GL_ATI_fragment_shader shader id\n");
 390         return 0;
 391     }
 392     GL_EXTCALL(glBindFragmentShaderATI(ret));
 393     checkGLcall("GL_EXTCALL(glBindFragmentShaderATI(ret))");
 394
 395     TRACE("glBeginFragmentShaderATI()\n");
 396     GL_EXTCALL(glBeginFragmentShaderATI());
 397     checkGLcall("GL_EXTCALL(glBeginFragmentShaderATI())");
 398
 399     /* Pass 1: Generate sampling instructions for perturbation maps */
 400     for (stage = 0; stage < gl_info->limits.textures; ++stage)
 401     {
 402         if(op[stage].cop == WINED3DTOP_DISABLE) break;
 403         if(op[stage].cop != WINED3DTOP_BUMPENVMAP &&
 404            op[stage].cop != WINED3DTOP_BUMPENVMAPLUMINANCE) continue;
 405
 406         TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, GL_SWIZZLE_STR_ATI)\n",
 407               stage, stage);
 408         GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 409                    GL_TEXTURE0_ARB + stage,
 410                    GL_SWIZZLE_STR_ATI));
 411         if(op[stage + 1].projected == proj_none) {
 412             swizzle = GL_SWIZZLE_STR_ATI;
 413         } else if(op[stage + 1].projected == proj_count4) {
 414             swizzle = GL_SWIZZLE_STQ_DQ_ATI;
 415         } else {
 416             swizzle = GL_SWIZZLE_STR_DR_ATI;
 417         }
 418         TRACE("glPassTexCoordATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, %s)\n",
 419               stage + 1, stage + 1, debug_swizzle(swizzle));
 420         GL_EXTCALL(glPassTexCoordATI(GL_REG_0_ATI + stage + 1,
 421                    GL_TEXTURE0_ARB + stage + 1,
 422                    swizzle));
 423     }
 424
 425     /* Pass 2: Generate perturbation calculations */
 426     for (stage = 0; stage < gl_info->limits.textures; ++stage)
 427     {
 428         GLuint argmodextra_x, argmodextra_y;
 429         struct color_fixup_desc fixup;
 430
 431         if(op[stage].cop == WINED3DTOP_DISABLE) break;
 432         if(op[stage].cop != WINED3DTOP_BUMPENVMAP &&
 433            op[stage].cop != WINED3DTOP_BUMPENVMAPLUMINANCE) continue;
 434
 435         fixup = op[stage].color_fixup;
 436         if (fixup.x_source != CHANNEL_SOURCE_X || fixup.y_source != CHANNEL_SOURCE_Y)
 437         {
 438             FIXME("Swizzles not implemented\n");
 439             argmodextra_x = GL_NONE;
 440             argmodextra_y = GL_NONE;
 441         }
 442         else
 443         {
 444             /* Nice thing, we get the color correction for free :-) */
 445             argmodextra_x = fixup.x_sign_fixup ? GL_2X_BIT_ATI | GL_BIAS_BIT_ATI : GL_NONE;
 446             argmodextra_y = fixup.y_sign_fixup ? GL_2X_BIT_ATI | GL_BIAS_BIT_ATI : GL_NONE;
 447         }
 448
 449         wrap_op3(gl_info, GL_DOT2_ADD_ATI, GL_REG_0_ATI + stage + 1, GL_RED_BIT_ATI, GL_NONE,
 450                  GL_REG_0_ATI + stage, GL_NONE, argmodextra_x,
 451                  ATI_FFP_CONST_BUMPMAT(stage), GL_NONE, GL_2X_BIT_ATI | GL_BIAS_BIT_ATI,
 452                  GL_REG_0_ATI + stage + 1, GL_RED, GL_NONE);
 453
 454         /* Don't use GL_DOT2_ADD_ATI here because we cannot configure it to read the blue and alpha
 455          * component of the bump matrix. Instead do this with two MADs:
 456          *
 457          * coord.a = tex.r * bump.b + coord.g
 458          * coord.g = tex.g * bump.a + coord.a
 459          *
 460          * The first instruction writes to alpha so it can be coissued with the above DOT2_ADD.
 461          * coord.a is unused. If the perturbed texture is projected, this was already handled
 462          * in the glPassTexCoordATI above.
 463          */
 464         wrap_op3(gl_info, GL_MAD_ATI, GL_REG_0_ATI + stage + 1, GL_ALPHA, GL_NONE,
 465                  GL_REG_0_ATI + stage, GL_RED, argmodextra_y,
 466                  ATI_FFP_CONST_BUMPMAT(stage), GL_BLUE, GL_NONE,
 467                  GL_REG_0_ATI + stage + 1, GL_GREEN, GL_NONE);
 468         wrap_op3(gl_info, GL_MAD_ATI, GL_REG_0_ATI + stage + 1, GL_GREEN_BIT_ATI, GL_NONE,
 469                  GL_REG_0_ATI + stage, GL_GREEN, argmodextra_y,
 470                  ATI_FFP_CONST_BUMPMAT(stage), GL_ALPHA, GL_NONE,
 471                  GL_REG_0_ATI + stage + 1, GL_ALPHA, GL_NONE);
 472     }
 473
 474     /* Pass 3: Generate sampling instructions for regular textures */
 475     for (stage = 0; stage < gl_info->limits.textures; ++stage)
 476     {
 477         if(op[stage].cop == WINED3DTOP_DISABLE) {
 478             break;
 479         }
 480
 481         if(op[stage].projected == proj_none) {
 482             swizzle = GL_SWIZZLE_STR_ATI;
 483         } else if(op[stage].projected == proj_count3) {
 484             swizzle = GL_SWIZZLE_STR_DR_ATI;
 485         } else {
 486             swizzle = GL_SWIZZLE_STQ_DQ_ATI;
 487         }
 488
 489         if((op[stage].carg0 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 490            (op[stage].carg1 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 491            (op[stage].carg2 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 492            (op[stage].aarg0 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 493            (op[stage].aarg1 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 494            (op[stage].aarg2 & WINED3DTA_SELECTMASK) == WINED3DTA_TEXTURE ||
 495             op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHA) {
 496
 497             if(stage > 0 &&
 498                (op[stage - 1].cop == WINED3DTOP_BUMPENVMAP ||
 499                 op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE)) {
 500                 TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_REG_%d_ATI, GL_SWIZZLE_STR_ATI)\n",
 501                       stage, stage);
 502                 GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 503                            GL_REG_0_ATI + stage,
 504                            GL_SWIZZLE_STR_ATI));
 505             } else {
 506                 TRACE("glSampleMapATI(GL_REG_%d_ATI, GL_TEXTURE_%d_ARB, %s)\n",
 507                     stage, stage, debug_swizzle(swizzle));
 508                 GL_EXTCALL(glSampleMapATI(GL_REG_0_ATI + stage,
 509                                         GL_TEXTURE0_ARB + stage,
 510                                         swizzle));
 511             }
 512         }
 513     }
 514
 515     /* Pass 4: Generate the arithmetic instructions */
 516     for (stage = 0; stage < MAX_TEXTURES; ++stage)
 517     {
 518         if (op[stage].cop == WINED3DTOP_DISABLE)
 519         {
 520             if (!stage)
 521             {
 522                 /* Handle complete texture disabling gracefully */
 523                 wrap_op1(gl_info, GL_MOV_ATI, GL_REG_0_ATI, GL_NONE, GL_NONE,
 524                          GL_PRIMARY_COLOR, GL_NONE, GL_NONE);
 525                 wrap_op1(gl_info, GL_MOV_ATI, GL_REG_0_ATI, GL_ALPHA, GL_NONE,
 526                          GL_PRIMARY_COLOR, GL_NONE, GL_NONE);
 527             }
 528             break;
 529         }
 530
 531         if(op[stage].dst == tempreg) {
 532             /* If we're writing to D3DTA_TEMP, but never reading from it we don't have to write there in the first place.
 533              * skip the entire stage, this saves some GPU time
 534              */
 535             if(tmparg == GL_NONE) continue;
 536
 537             dstreg = tmparg;
 538         } else {
 539             dstreg = GL_REG_0_ATI;
 540         }
 541
 542         arg0 = register_for_arg(op[stage].carg0, gl_info, stage, &argmod0, &rep0, tmparg);
 543         arg1 = register_for_arg(op[stage].carg1, gl_info, stage, &argmod1, &rep1, tmparg);
 544         arg2 = register_for_arg(op[stage].carg2, gl_info, stage, &argmod2, &rep2, tmparg);
 545         dstmod = GL_NONE;
 546         argmodextra = GL_NONE;
 547         extrarg = GL_NONE;
 548
 549         switch(op[stage].cop) {
 550             case WINED3DTOP_SELECTARG2:
 551                 arg1 = arg2;
 552                 argmod1 = argmod2;
 553                 rep1 = rep2;
 554             case WINED3DTOP_SELECTARG1:
 555                 wrap_op1(gl_info, GL_MOV_ATI, dstreg, GL_NONE, GL_NONE,
 556                          arg1, rep1, argmod1);
 557                 break;
 558
 559             case WINED3DTOP_MODULATE4X:
 560                 if(dstmod == GL_NONE) dstmod = GL_4X_BIT_ATI;
 561             case WINED3DTOP_MODULATE2X:
 562                 if(dstmod == GL_NONE) dstmod = GL_2X_BIT_ATI;
 563                 dstmod |= GL_SATURATE_BIT_ATI;
 564             case WINED3DTOP_MODULATE:
 565                 wrap_op2(gl_info, GL_MUL_ATI, dstreg, GL_NONE, dstmod,
 566                          arg1, rep1, argmod1,
 567                          arg2, rep2, argmod2);
 568                 break;
 569
 570             case WINED3DTOP_ADDSIGNED2X:
 571                 dstmod = GL_2X_BIT_ATI;
 572             case WINED3DTOP_ADDSIGNED:
 573                 argmodextra = GL_BIAS_BIT_ATI;
 574             case WINED3DTOP_ADD:
 575                 dstmod |= GL_SATURATE_BIT_ATI;
 576                 wrap_op2(gl_info, GL_ADD_ATI, GL_REG_0_ATI, GL_NONE, dstmod,
 577                          arg1, rep1, argmod1,
 578                          arg2, rep2, argmodextra | argmod2);
 579                 break;
 580
 581             case WINED3DTOP_SUBTRACT:
 582                 dstmod |= GL_SATURATE_BIT_ATI;
 583                 wrap_op2(gl_info, GL_SUB_ATI, dstreg, GL_NONE, dstmod,
 584                          arg1, rep1, argmod1,
 585                          arg2, rep2, argmod2);
 586                 break;
 587
 588             case WINED3DTOP_ADDSMOOTH:
 589                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 590                 /* Dst = arg1 + * arg2(1 -arg 1)
 591                  *     = arg2 * (1 - arg1) + arg1
 592                  */
 593                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 594                          arg2, rep2, argmod2,
 595                          arg1, rep1, argmodextra,
 596                          arg1, rep1, argmod1);
 597                 break;
 598
 599             case WINED3DTOP_BLENDCURRENTALPHA:
 600                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_CURRENT, gl_info, stage, NULL, NULL, -1);
 601             case WINED3DTOP_BLENDFACTORALPHA:
 602                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TFACTOR, gl_info, stage, NULL, NULL, -1);
 603             case WINED3DTOP_BLENDTEXTUREALPHA:
 604                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 605             case WINED3DTOP_BLENDDIFFUSEALPHA:
 606                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_DIFFUSE, gl_info, stage, NULL, NULL, -1);
 607                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_NONE, GL_NONE,
 608                          extrarg, GL_ALPHA, GL_NONE,
 609                          arg1, rep1, argmod1,
 610                          arg2, rep2, argmod2);
 611                 break;
 612
 613             case WINED3DTOP_BLENDTEXTUREALPHAPM:
 614                 arg0 = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 615                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_NONE,
 616                          arg2, rep2,  argmod2,
 617                          arg0, GL_ALPHA, GL_COMP_BIT_ATI,
 618                          arg1, rep1,  argmod1);
 619                 break;
 620
 621             /* D3DTOP_PREMODULATE ???? */
 622
 623             case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
 624                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 625             case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
 626                 if(!argmodextra) argmodextra = argmod1;
 627                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 628                          arg2, rep2,  argmod2,
 629                          arg1, GL_ALPHA, argmodextra,
 630                          arg1, rep1,  argmod1);
 631                 break;
 632
 633             case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
 634                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 635             case WINED3DTOP_MODULATECOLOR_ADDALPHA:
 636                 if(!argmodextra) argmodextra = argmod1;
 637                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 638                          arg2, rep2,  argmod2,
 639                          arg1, rep1,  argmodextra,
 640                          arg1, GL_ALPHA, argmod1);
 641                 break;
 642
 643             case WINED3DTOP_DOTPRODUCT3:
 644                 wrap_op2(gl_info, GL_DOT3_ATI, dstreg, GL_NONE, GL_4X_BIT_ATI | GL_SATURATE_BIT_ATI,
 645                          arg1, rep1, argmod1 | GL_BIAS_BIT_ATI,
 646                          arg2, rep2, argmod2 | GL_BIAS_BIT_ATI);
 647                 break;
 648
 649             case WINED3DTOP_MULTIPLYADD:
 650                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_NONE, GL_SATURATE_BIT_ATI,
 651                          arg1, rep1, argmod1,
 652                          arg2, rep2, argmod2,
 653                          arg0, rep0, argmod0);
 654                 break;
 655
 656             case WINED3DTOP_LERP:
 657                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_NONE, GL_NONE,
 658                          arg0, rep0, argmod0,
 659                          arg1, rep1, argmod1,
 660                          arg2, rep2, argmod2);
 661                 break;
 662
 663             case WINED3DTOP_BUMPENVMAP:
 664             case WINED3DTOP_BUMPENVMAPLUMINANCE:
 665                 /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
 666                 break;
 667
 668             default: FIXME("Unhandled color operation %d on stage %d\n", op[stage].cop, stage);
 669         }
 670
 671         arg0 = register_for_arg(op[stage].aarg0, gl_info, stage, &argmod0, NULL, tmparg);
 672         arg1 = register_for_arg(op[stage].aarg1, gl_info, stage, &argmod1, NULL, tmparg);
 673         arg2 = register_for_arg(op[stage].aarg2, gl_info, stage, &argmod2, NULL, tmparg);
 674         dstmod = GL_NONE;
 675         argmodextra = GL_NONE;
 676         extrarg = GL_NONE;
 677
 678         switch(op[stage].aop) {
 679             case WINED3DTOP_DISABLE:
 680                 /* Get the primary color to the output if on stage 0, otherwise leave register 0 untouched */
 681                 if (!stage)
 682                 {
 683                     wrap_op1(gl_info, GL_MOV_ATI, GL_REG_0_ATI, GL_ALPHA, GL_NONE,
 684                              GL_PRIMARY_COLOR, GL_NONE, GL_NONE);
 685                 }
 686                 break;
 687
 688             case WINED3DTOP_SELECTARG2:
 689                 arg1 = arg2;
 690                 argmod1 = argmod2;
 691             case WINED3DTOP_SELECTARG1:
 692                 wrap_op1(gl_info, GL_MOV_ATI, dstreg, GL_ALPHA, GL_NONE,
 693                          arg1, GL_NONE, argmod1);
 694                 break;
 695
 696             case WINED3DTOP_MODULATE4X:
 697                 if(dstmod == GL_NONE) dstmod = GL_4X_BIT_ATI;
 698             case WINED3DTOP_MODULATE2X:
 699                 if(dstmod == GL_NONE) dstmod = GL_2X_BIT_ATI;
 700                 dstmod |= GL_SATURATE_BIT_ATI;
 701             case WINED3DTOP_MODULATE:
 702                 wrap_op2(gl_info, GL_MUL_ATI, dstreg, GL_ALPHA, dstmod,
 703                          arg1, GL_NONE, argmod1,
 704                          arg2, GL_NONE, argmod2);
 705                 break;
 706
 707             case WINED3DTOP_ADDSIGNED2X:
 708                 dstmod = GL_2X_BIT_ATI;
 709             case WINED3DTOP_ADDSIGNED:
 710                 argmodextra = GL_BIAS_BIT_ATI;
 711             case WINED3DTOP_ADD:
 712                 dstmod |= GL_SATURATE_BIT_ATI;
 713                 wrap_op2(gl_info, GL_ADD_ATI, dstreg, GL_ALPHA, dstmod,
 714                          arg1, GL_NONE, argmod1,
 715                          arg2, GL_NONE, argmodextra | argmod2);
 716                 break;
 717
 718             case WINED3DTOP_SUBTRACT:
 719                 dstmod |= GL_SATURATE_BIT_ATI;
 720                 wrap_op2(gl_info, GL_SUB_ATI, dstreg, GL_ALPHA, dstmod,
 721                          arg1, GL_NONE, argmod1,
 722                          arg2, GL_NONE, argmod2);
 723                 break;
 724
 725             case WINED3DTOP_ADDSMOOTH:
 726                 argmodextra = argmod1 & GL_COMP_BIT_ATI ? argmod1 & ~GL_COMP_BIT_ATI : argmod1 | GL_COMP_BIT_ATI;
 727                 /* Dst = arg1 + * arg2(1 -arg 1)
 728                  *     = arg2 * (1 - arg1) + arg1
 729                  */
 730                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_ALPHA, GL_SATURATE_BIT_ATI,
 731                          arg2, GL_NONE, argmod2,
 732                          arg1, GL_NONE, argmodextra,
 733                          arg1, GL_NONE, argmod1);
 734                 break;
 735
 736             case WINED3DTOP_BLENDCURRENTALPHA:
 737                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_CURRENT, gl_info, stage, NULL, NULL, -1);
 738             case WINED3DTOP_BLENDFACTORALPHA:
 739                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TFACTOR, gl_info, stage, NULL, NULL, -1);
 740             case WINED3DTOP_BLENDTEXTUREALPHA:
 741                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 742             case WINED3DTOP_BLENDDIFFUSEALPHA:
 743                 if(extrarg == GL_NONE) extrarg = register_for_arg(WINED3DTA_DIFFUSE, gl_info, stage, NULL, NULL, -1);
 744                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_ALPHA, GL_NONE,
 745                          extrarg, GL_ALPHA, GL_NONE,
 746                          arg1, GL_NONE, argmod1,
 747                          arg2, GL_NONE, argmod2);
 748                 break;
 749
 750             case WINED3DTOP_BLENDTEXTUREALPHAPM:
 751                 arg0 = register_for_arg(WINED3DTA_TEXTURE, gl_info, stage, NULL, NULL, -1);
 752                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_ALPHA, GL_NONE,
 753                          arg2, GL_NONE,  argmod2,
 754                          arg0, GL_ALPHA, GL_COMP_BIT_ATI,
 755                          arg1, GL_NONE,  argmod1);
 756                 break;
 757
 758             /* D3DTOP_PREMODULATE ???? */
 759
 760             case WINED3DTOP_DOTPRODUCT3:
 761                 wrap_op2(gl_info, GL_DOT3_ATI, dstreg, GL_ALPHA, GL_4X_BIT_ATI | GL_SATURATE_BIT_ATI,
 762                          arg1, GL_NONE, argmod1 | GL_BIAS_BIT_ATI,
 763                          arg2, GL_NONE, argmod2 | GL_BIAS_BIT_ATI);
 764                 break;
 765
 766             case WINED3DTOP_MULTIPLYADD:
 767                 wrap_op3(gl_info, GL_MAD_ATI, dstreg, GL_ALPHA, GL_SATURATE_BIT_ATI,
 768                          arg1, GL_NONE, argmod1,
 769                          arg2, GL_NONE, argmod2,
 770                          arg0, GL_NONE, argmod0);
 771                 break;
 772
 773             case WINED3DTOP_LERP:
 774                 wrap_op3(gl_info, GL_LERP_ATI, dstreg, GL_ALPHA, GL_SATURATE_BIT_ATI,
 775                          arg1, GL_NONE, argmod1,
 776                          arg2, GL_NONE, argmod2,
 777                          arg0, GL_NONE, argmod0);
 778                 break;
 779
 780             case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
 781             case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
 782             case WINED3DTOP_MODULATECOLOR_ADDALPHA:
 783             case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
 784             case WINED3DTOP_BUMPENVMAP:
 785             case WINED3DTOP_BUMPENVMAPLUMINANCE:
 786                 ERR("Application uses an invalid alpha operation\n");
 787                 break;
 788
 789             default: FIXME("Unhandled alpha operation %d on stage %d\n", op[stage].aop, stage);
 790         }
 791     }
 792
 793     TRACE("glEndFragmentShaderATI()\n");
 794     GL_EXTCALL(glEndFragmentShaderATI());
 795     checkGLcall("GL_EXTCALL(glEndFragmentShaderATI())");
 796     return ret;
 797 }
 798
 799 static void set_tex_op_atifs(DWORD state, struct wined3d_stateblock *stateblock, struct wined3d_context *context)
 800 {
 801     const struct wined3d_gl_info *gl_info = context->gl_info;
 802     IWineD3DDeviceImpl *This = stateblock->device;
 803     const struct atifs_ffp_desc *desc;
 804     struct ffp_frag_settings     settings;
 805     struct atifs_private_data   *priv = This->fragment_priv;
 806     DWORD mapped_stage;
 807     unsigned int i;
 808
 809     gen_ffp_frag_op(stateblock, &settings, TRUE);
 810     desc = (const struct atifs_ffp_desc *)find_ffp_frag_shader(&priv->fragment_shaders, &settings);
 811     if(!desc) {
 812         struct atifs_ffp_desc *new_desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*new_desc));
 813         if (!new_desc)
 814         {
 815             ERR("Out of memory\n");
 816             return;
 817         }
 818         new_desc->num_textures_used = 0;
 819         for (i = 0; i < gl_info->limits.texture_stages; ++i)
 820         {
 821             if(settings.op[i].cop == WINED3DTOP_DISABLE) break;
 822             new_desc->num_textures_used = i;
 823         }
 824
 825         memcpy(&new_desc->parent.settings, &settings, sizeof(settings));
 826         new_desc->shader = gen_ati_shader(settings.op, gl_info);
 827         add_ffp_frag_shader(&priv->fragment_shaders, &new_desc->parent);
 828         TRACE("Allocated fixed function replacement shader descriptor %p\n", new_desc);
 829         desc = new_desc;
 830     }
 831
 832     /* GL_ATI_fragment_shader depends on the GL_TEXTURE_xD enable settings. Update the texture stages
 833      * used by this shader
 834      */
 835     for(i = 0; i < desc->num_textures_used; i++) {
 836         mapped_stage = This->texUnitMap[i];
 837         if (mapped_stage != WINED3D_UNMAPPED_STAGE)
 838         {
 839             GL_EXTCALL(glActiveTextureARB(GL_TEXTURE0_ARB + mapped_stage));
 840             checkGLcall("glActiveTextureARB");
 841             texture_activate_dimensions(stateblock->state.textures[i], gl_info);
 842         }
 843     }
 844
 845     GL_EXTCALL(glBindFragmentShaderATI(desc->shader));
 846 }
 847
 848 static void state_texfactor_atifs(DWORD state, struct wined3d_stateblock *stateblock, struct wined3d_context *context)
 849 {
 850     const struct wined3d_gl_info *gl_info = context->gl_info;
 851     float col[4];
 852     D3DCOLORTOGLFLOAT4(stateblock->state.render_states[WINED3DRS_TEXTUREFACTOR], col);
 853
 854     GL_EXTCALL(glSetFragmentShaderConstantATI(ATI_FFP_CONST_TFACTOR, col));
 855     checkGLcall("glSetFragmentShaderConstantATI(ATI_FFP_CONST_TFACTOR, col)");
 856 }
 857
 858 static void set_bumpmat(DWORD state, struct wined3d_stateblock *stateblock, struct wined3d_context *context)
 859 {
 860     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
 861     const struct wined3d_gl_info *gl_info = context->gl_info;
 862     float mat[2][2];
 863
 864     mat[0][0] = *((float *)&stateblock->state.texture_states[stage][WINED3DTSS_BUMPENVMAT00]);
 865     mat[1][0] = *((float *)&stateblock->state.texture_states[stage][WINED3DTSS_BUMPENVMAT01]);
 866     mat[0][1] = *((float *)&stateblock->state.texture_states[stage][WINED3DTSS_BUMPENVMAT10]);
 867     mat[1][1] = *((float *)&stateblock->state.texture_states[stage][WINED3DTSS_BUMPENVMAT11]);
 868     /* GL_ATI_fragment_shader allows only constants from 0.0 to 1.0, but the bumpmat
 869      * constants can be in any range. While they should stay between [-1.0 and 1.0] because
 870      * Shader Model 1.x pixel shaders are clamped to that range negative values are used occasionally,
 871      * for example by our d3d9 test. So to get negative values scale -1;1 to 0;1 and undo that in the
 872      * shader(it is free). This might potentially reduce precision. However, if the hardware does
 873      * support proper floats it shouldn't, and if it doesn't we can't get anything better anyway
 874      */
 875     mat[0][0] = (mat[0][0] + 1.0f) * 0.5f;
 876     mat[1][0] = (mat[1][0] + 1.0f) * 0.5f;
 877     mat[0][1] = (mat[0][1] + 1.0f) * 0.5f;
 878     mat[1][1] = (mat[1][1] + 1.0f) * 0.5f;
 879     GL_EXTCALL(glSetFragmentShaderConstantATI(ATI_FFP_CONST_BUMPMAT(stage), (float *) mat));
 880     checkGLcall("glSetFragmentShaderConstantATI(ATI_FFP_CONST_BUMPMAT(stage), mat)");
 881 }
 882
 883 static void textransform(DWORD state, struct wined3d_stateblock *stateblock, struct wined3d_context *context)
 884 {
 885     if (!isStateDirty(context, STATE_PIXELSHADER))
 886         set_tex_op_atifs(state, stateblock, context);
 887 }
 888
 889 static void atifs_apply_pixelshader(DWORD state_id,
 890         struct wined3d_stateblock *stateblock, struct wined3d_context *context)
 891 {
 892     const struct wined3d_state *state = &stateblock->state;
 893     IWineD3DDeviceImpl *device = stateblock->device;
 894     BOOL use_vshader = use_vs(state);
 895
 896     context->last_was_pshader = use_ps(state);
 897     /* The ATIFS code does not support pixel shaders currently, but we have to provide a state handler
 898      * to call shader_select to select a vertex shader if one is applied because the vertex shader state
 899      * may defer calling the shader backend if the pshader state is dirty.
 900      *
 901      * In theory the application should not be able to mark the pixel shader dirty because it cannot
 902      * create a shader, and thus has no way to set the state to something != NULL. However, a different
 903      * pipeline part may link a different state to its pixelshader handler, thus a pshader state exists
 904      * and can be dirtified. Also the pshader is always dirtified at startup, and blitting disables all
 905      * shaders and dirtifies all shader states. If atifs can deal with this it keeps the rest of the code
 906      * simpler.
 907      */
 908     if(!isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
 909         device->shader_backend->shader_select(context, FALSE, use_vshader);
 910
 911         if (!isStateDirty(context, STATE_VERTEXSHADERCONSTANT) && use_vshader)
 912             stateblock_apply_state(STATE_VERTEXSHADERCONSTANT, stateblock, context);
 913     }
 914 }
 915
 916 static void atifs_srgbwriteenable(DWORD state, struct wined3d_stateblock *stateblock, struct wined3d_context *context)
 917 {
 918     if (stateblock->state.render_states[WINED3DRS_SRGBWRITEENABLE])
 919         WARN("sRGB writes are not supported by this fragment pipe.\n");
 920 }
 921
 922 static const struct StateEntryTemplate atifs_fragmentstate_template[] = {
 923     {STATE_RENDER(WINED3DRS_TEXTUREFACTOR),               { STATE_RENDER(WINED3DRS_TEXTUREFACTOR),              state_texfactor_atifs   }, WINED3D_GL_EXT_NONE             },
 924     {STATE_RENDER(WINED3DRS_FOGCOLOR),                    { STATE_RENDER(WINED3DRS_FOGCOLOR),                   state_fogcolor          }, WINED3D_GL_EXT_NONE             },
 925     {STATE_RENDER(WINED3DRS_FOGDENSITY),                  { STATE_RENDER(WINED3DRS_FOGDENSITY),                 state_fogdensity        }, WINED3D_GL_EXT_NONE             },
 926     {STATE_RENDER(WINED3DRS_FOGENABLE),                   { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_fog_fragpart      }, WINED3D_GL_EXT_NONE             },
 927     {STATE_RENDER(WINED3DRS_FOGTABLEMODE),                { STATE_RENDER(WINED3DRS_FOGENABLE),                  NULL                    }, WINED3D_GL_EXT_NONE             },
 928     {STATE_RENDER(WINED3DRS_FOGVERTEXMODE),               { STATE_RENDER(WINED3DRS_FOGENABLE),                  NULL                    }, WINED3D_GL_EXT_NONE             },
 929     {STATE_RENDER(WINED3DRS_FOGSTART),                    { STATE_RENDER(WINED3DRS_FOGSTART),                   state_fogstartend       }, WINED3D_GL_EXT_NONE             },
 930     {STATE_RENDER(WINED3DRS_FOGEND),                      { STATE_RENDER(WINED3DRS_FOGSTART),                   NULL                    }, WINED3D_GL_EXT_NONE             },
 931     {STATE_RENDER(WINED3DRS_SRGBWRITEENABLE),             { STATE_RENDER(WINED3DRS_SRGBWRITEENABLE),            atifs_srgbwriteenable   }, WINED3D_GL_EXT_NONE             },
 932     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          set_tex_op_atifs        }, WINED3D_GL_EXT_NONE             },
 933     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 934     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 935     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 936     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 937     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 938     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 939     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 940     {STATE_TEXTURESTAGE(0, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 941     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
 942     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 943     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 944     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 945     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 946     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 947     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 948     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 949     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 950     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 951     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 952     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 953     {STATE_TEXTURESTAGE(1, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 954     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
 955     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 956     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 957     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 958     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 959     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 960     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 961     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 962     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 963     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 964     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 965     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 966     {STATE_TEXTURESTAGE(2, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 967     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
 968     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 969     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 970     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 971     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 972     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 973     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 974     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 975     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 976     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 977     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 978     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 979     {STATE_TEXTURESTAGE(3, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 980     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
 981     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 982     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 983     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 984     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 985     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 986     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 987     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 988     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 989     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 990     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 991     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 992     {STATE_TEXTURESTAGE(4, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 993     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
 994     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 995     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 996     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
 997     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 998     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
 999     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1000     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1001     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1002     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1003     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1004     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1005     {STATE_TEXTURESTAGE(5, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1006     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1007     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1008     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1009     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1010     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1011     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1012     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1013     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1014     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1015     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1016     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1017     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1018     {STATE_TEXTURESTAGE(6, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1019     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1020     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1021     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1022     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1023     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLOROP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1024     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1025     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1026     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1027     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAOP),           { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1028     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG1),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1029     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG2),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1030     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG0),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1031     {STATE_TEXTURESTAGE(7, WINED3DTSS_RESULTARG),         { STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),          NULL                    }, WINED3D_GL_EXT_NONE             },
1032     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat             }, WINED3D_GL_EXT_NONE             },
1033     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1034     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1035     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     NULL                    }, WINED3D_GL_EXT_NONE             },
1036     { STATE_SAMPLER(0),                                   { STATE_SAMPLER(0),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1037     { STATE_SAMPLER(1),                                   { STATE_SAMPLER(1),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1038     { STATE_SAMPLER(2),                                   { STATE_SAMPLER(2),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1039     { STATE_SAMPLER(3),                                   { STATE_SAMPLER(3),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1040     { STATE_SAMPLER(4),                                   { STATE_SAMPLER(4),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1041     { STATE_SAMPLER(5),                                   { STATE_SAMPLER(5),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1042     { STATE_SAMPLER(6),                                   { STATE_SAMPLER(6),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1043     { STATE_SAMPLER(7),                                   { STATE_SAMPLER(7),                                   sampler_texdim          }, WINED3D_GL_EXT_NONE             },
1044     {STATE_TEXTURESTAGE(0,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(0, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1045     {STATE_TEXTURESTAGE(1,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(1, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1046     {STATE_TEXTURESTAGE(2,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(2, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1047     {STATE_TEXTURESTAGE(3,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(3, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1048     {STATE_TEXTURESTAGE(4,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(4, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1049     {STATE_TEXTURESTAGE(5,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(5, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1050     {STATE_TEXTURESTAGE(6,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(6, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1051     {STATE_TEXTURESTAGE(7,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(7, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, WINED3D_GL_EXT_NONE             },
1052     {STATE_PIXELSHADER,                                   { STATE_PIXELSHADER,                                  atifs_apply_pixelshader }, WINED3D_GL_EXT_NONE             },
1053     {0 /* Terminate */,                                   { 0,                                                  0                       }, WINED3D_GL_EXT_NONE             },
1054 };
1055
1056 /* Context activation is done by the caller. */
1057 static void atifs_enable(BOOL enable)
1058 {
1059     ENTER_GL();
1060     if(enable) {
1061         glEnable(GL_FRAGMENT_SHADER_ATI);
1062         checkGLcall("glEnable(GL_FRAGMENT_SHADER_ATI)");
1063     } else {
1064         glDisable(GL_FRAGMENT_SHADER_ATI);
1065         checkGLcall("glDisable(GL_FRAGMENT_SHADER_ATI)");
1066     }
1067     LEAVE_GL();
1068 }
1069
1070 static void atifs_get_caps(const struct wined3d_gl_info *gl_info, struct fragment_caps *caps)
1071 {
1072     caps->PrimitiveMiscCaps = WINED3DPMISCCAPS_TSSARGTEMP;
1073     caps->TextureOpCaps =  WINED3DTEXOPCAPS_DISABLE                     |
1074                            WINED3DTEXOPCAPS_SELECTARG1                  |
1075                            WINED3DTEXOPCAPS_SELECTARG2                  |
1076                            WINED3DTEXOPCAPS_MODULATE4X                  |
1077                            WINED3DTEXOPCAPS_MODULATE2X                  |
1078                            WINED3DTEXOPCAPS_MODULATE                    |
1079                            WINED3DTEXOPCAPS_ADDSIGNED2X                 |
1080                            WINED3DTEXOPCAPS_ADDSIGNED                   |
1081                            WINED3DTEXOPCAPS_ADD                         |
1082                            WINED3DTEXOPCAPS_SUBTRACT                    |
1083                            WINED3DTEXOPCAPS_ADDSMOOTH                   |
1084                            WINED3DTEXOPCAPS_BLENDCURRENTALPHA           |
1085                            WINED3DTEXOPCAPS_BLENDFACTORALPHA            |
1086                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHA           |
1087                            WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA           |
1088                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM         |
1089                            WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR      |
1090                            WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA      |
1091                            WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA   |
1092                            WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR   |
1093                            WINED3DTEXOPCAPS_DOTPRODUCT3                 |
1094                            WINED3DTEXOPCAPS_MULTIPLYADD                 |
1095                            WINED3DTEXOPCAPS_LERP                        |
1096                            WINED3DTEXOPCAPS_BUMPENVMAP;
1097
1098     /* TODO: Implement WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE
1099     and WINED3DTEXOPCAPS_PREMODULATE */
1100
1101     /* GL_ATI_fragment_shader always supports 6 textures, which was the limit on r200 cards
1102      * which this extension is exclusively focused on(later cards have GL_ARB_fragment_program).
1103      * If the current card has more than 8 fixed function textures in OpenGL's regular fixed
1104      * function pipeline then the ATI_fragment_shader backend imposes a stricter limit. This
1105      * shouldn't be too hard since Nvidia cards have a limit of 4 textures with the default ffp
1106      * pipeline, and almost all games are happy with that. We can however support up to 8
1107      * texture stages because we have a 2nd pass limit of 8 instructions, and per stage we use
1108      * only 1 instruction.
1109      *
1110      * The proper fix for this is not to use GL_ATI_fragment_shader on cards newer than the
1111      * r200 series and use an ARB or GLSL shader instead
1112      */
1113     caps->MaxTextureBlendStages   = 8;
1114     caps->MaxSimultaneousTextures = 6;
1115 }
1116
1117 static HRESULT atifs_alloc(IWineD3DDeviceImpl *device)
1118 {
1119     struct atifs_private_data *priv;
1120
1121     device->fragment_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct atifs_private_data));
1122     if (!device->fragment_priv)
1123     {
1124         ERR("Out of memory\n");
1125         return E_OUTOFMEMORY;
1126     }
1127     priv = device->fragment_priv;
1128     if (wine_rb_init(&priv->fragment_shaders, &wined3d_ffp_frag_program_rb_functions) == -1)
1129     {
1130         ERR("Failed to initialize rbtree.\n");
1131         HeapFree(GetProcessHeap(), 0, device->fragment_priv);
1132         return E_OUTOFMEMORY;
1133     }
1134     return WINED3D_OK;
1135 }
1136
1137 /* Context activation is done by the caller. */
1138 static void atifs_free_ffpshader(struct wine_rb_entry *entry, void *context)
1139 {
1140     IWineD3DDeviceImpl *This = context;
1141     const struct wined3d_gl_info *gl_info = &This->adapter->gl_info;
1142     struct atifs_ffp_desc *entry_ati = WINE_RB_ENTRY_VALUE(entry, struct atifs_ffp_desc, parent.entry);
1143
1144     ENTER_GL();
1145     GL_EXTCALL(glDeleteFragmentShaderATI(entry_ati->shader));
1146     checkGLcall("glDeleteFragmentShaderATI(entry->shader)");
1147     HeapFree(GetProcessHeap(), 0, entry_ati);
1148     LEAVE_GL();
1149 }
1150
1151 /* Context activation is done by the caller. */
1152 static void atifs_free(IWineD3DDeviceImpl *device)
1153 {
1154     struct atifs_private_data *priv = device->fragment_priv;
1155
1156     wine_rb_destroy(&priv->fragment_shaders, atifs_free_ffpshader, device);
1157
1158     HeapFree(GetProcessHeap(), 0, priv);
1159     device->fragment_priv = NULL;
1160 }
1161
1162 static BOOL atifs_color_fixup_supported(struct color_fixup_desc fixup)
1163 {
1164     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1165     {
1166         TRACE("Checking support for fixup:\n");
1167         dump_color_fixup_desc(fixup);
1168     }
1169
1170     /* We only support sign fixup of the first two channels. */
1171     if (fixup.x_source == CHANNEL_SOURCE_X && fixup.y_source == CHANNEL_SOURCE_Y
1172             && fixup.z_source == CHANNEL_SOURCE_Z && fixup.w_source == CHANNEL_SOURCE_W
1173             && !fixup.z_sign_fixup && !fixup.w_sign_fixup)
1174     {
1175         TRACE("[OK]\n");
1176         return TRUE;
1177     }
1178
1179     TRACE("[FAILED]\n");
1180     return FALSE;
1181 }
1182
1183 const struct fragment_pipeline atifs_fragment_pipeline = {
1184     atifs_enable,
1185     atifs_get_caps,
1186     atifs_alloc,
1187     atifs_free,
1188     atifs_color_fixup_supported,
1189     atifs_fragmentstate_template,
1190     TRUE /* We can disable projected textures */
1191 };